MindAlign: Decoding Inner Speech from fMRI Signals via Multimodal Embedding Alignment under Limited Data

Researchers introduce MindAlign, a two-stage framework that decodes inner speech from fMRI brain signals by aligning neural activity with semantic embeddings, then using a frozen language model for text generation. The approach demonstrates improved performance over existing methods and shows that semantic-to-language mappings can generalize across subjects, advancing scalable brain-to-text decoding technology.

MindAlign represents a significant advancement in neurotechnology by addressing a persistent challenge in brain-computer interfaces: translating non-invasive brain signals into coherent language without subject-specific model retraining. The framework's two-stage architecture separates neural alignment from language generation, enabling modularity that existing task-specific approaches lack. This decoupling is crucial for practical deployment since it eliminates the need to fine-tune language models for each new participant, reducing computational overhead and implementation complexity.

The research builds on growing interest in brain-signal decoding, where previous systems struggled with limited training data, high inter-subject variability, and scalability constraints. By leveraging multimodal embedding spaces and frozen language models, MindAlign sidesteps these bottlenecks. The finding that semantic-to-language projections generalize across subjects suggests that despite individual neural differences, underlying semantic representations follow transferable patterns—a breakthrough for reducing per-subject training requirements.

For the neurotechnology and AI sectors, this work validates a modular approach to brain-to-text systems that could accelerate development of assistive technologies for locked-in patients and communication disorders. The ability to extract semantic content independent of visual priors demonstrates neural signals carry rich information beyond stimulus-driven responses, expanding potential applications beyond simple signal decoding.

Future developments may focus on real-time performance, non-fMRI brain signal compatibility (EEG, ECoG), and clinical translation. As brain-computer interfaces move toward commercial deployment, scalable frameworks like MindAlign become increasingly valuable for reducing individualization overhead while maintaining accuracy.

• →MindAlign decodes inner speech from fMRI without requiring language model fine-tuning for each new subject, improving scalability.
• →The two-stage approach separates neural-semantic alignment from language generation, enabling modular and transferable components.
• →Semantic-to-language projections generalize across subjects, suggesting shared semantic representations despite individual neural variability.
• →The framework outperforms existing fMRI-only and random baselines in open-ended text generation tasks.
• →Neural signals modulate semantic content independently of visual priors, indicating richer information content for brain-to-text applications.